You can win $20K in cash prizes by building a demo, training a model, or creating a dataset with Beam by @reflection_ai!

Prizes are split across 3 winners that we’ll choose after the challenge ends on Monday, November 2nd.

Plus, we’re giving $25 hub credits to the first 400 to take part, just find the link in the page below.

Rules: Build something cool with beam and drop it on the discussion! Anything that might be helpful to the world or a cool demo is allowed:

• You could make a demo that uses the Beam model on Spaces
• You could quantize it for your favorite hardware
• You could generate a synthetic dataset for your favorite problem
• You could build an RL environment with or for Beam
• You could benchmark Beam and share the results.
• You could surprise us!

If you build something with Beam, then share it on the Hub to enter the challenge. We’ll collect all the links, select the best, and reach out to winners.

https://huggingface.co/reflection/Beam-1.0-501B-A23B/discussions/5…

reflection/Beam-1.0-501B-A23B · Build with Beam - $20K in prizes

Source: https://huggingface.co/reflection/Beam-1.0-501B-A23B/discussions/5 Reflection org

13 days ago

•

edited 2 days ago We are updating our challenge to offer a total of $20K in prizes - split across 3 winners that we’ll choose after the challenge ends. You can choose between cash or HF credits. Deadline to submit: Monday, November 2nd.

Plus, we’re giving $25 hub credits to the first 400 people that claim with this link - https://huggingface.co/coupons/claim/HuggingFacexReflection-MiniHack

The rules: Build something cool with beam and drop it here! Anything that might be helpful to the world or a cool demo is allowed:

You could make a demo that uses the Beam model on Spaces You could quantize it for your favorite hardware You could generate a synthetic dataset for your favorite problem You could build an RL environment with or for Beam You could benchmark Beam and share the results. You could surprise us!

If you build something with Beam, then share it on the Hub to enter the challenge. We’ll collect all the links, select the best, and reach out to winners.

Share what you build here!

Update on judging criteria and prize distribution

Submission deadline: November 2, 2026, at 11:59PM PT

Reflection will appoint three (3) judges to evaluate eligible Submissions based on the following criteria, each weighted equally:

• Creativity and originality: 25%
• Technical execution: 25%
• Usefulness or potential real-world impact: 25%
• Innovative or effective use of Beam: 25%

The prizes will be allocated as follows:

• 1st place: $9,000
• 2nd place: $7,000
• 3rd place: $4,000

Total: $20,000 USD

The Official Rules for the Build with Beam competition have been publishedhere. Please review them carefully, as they govern participation in the competition, including eligibility, submissions, judging, and prizes. By submitting an entry, participants agree to be bound by the Official Rules. All potential winners will be required to affirmatively agree to be bound by the Official Rules before receiving a prize.

Is there any Transformers/PyTorch training code available? I dont see it on your github.

🚀 Excited to shareBeam Safety Benchmark: our contribution to the Reflection Beam community challenge!

🛡️ We’ve built an interactive Hugging Face Space dedicated to exploring the safety and reliability of Reflection’s Beam model.

As part of our ongoing work on trustworthy AI and model auditing, we’re excited to put Beam under the microscope and help make safety evaluation more accessible, transparent, and reproducible. 🔗 Try it here:https://huggingface.co/spaces/SushantGautam/BeamSafetyBenchmark

Built with curiosity, open science, and the belief that powerful AI deserves rigorous evaluation. Thanks to @Reflection and @huggingface for supporting the open AI community!

#Beam #AISafety #OpenScience@kelkalot

📡Beam Safety Benchmark— auditing Beam with SimpleAudit, live-updating results

I built an AI-safety audit ofBeamusingSimpleAudit(Simula’s open auditing framework) and published the results as a live, self-contained report that updates as new runs land.

🔗 Live results (static HF Space):https://sushantgautam-beamsafetybenchmark.static.hf.space/index.html 📦 Repo:https://github.com/SushantGautam/ReflectionDemo

How Beam was run:

• Target =Beam-1.0-501B-A23Bvia Reflection’s official OpenAI-compatible API (api\.reflection\.ai), served serverlessly — no local GPUs.
• Beam always reasons; there is no “off” switch, so I ran it in its minimum thinking mode (reasoning\_effort=low) to keep reasoning cost down while still exercising multi-turn conversations (up to 5 turns).

Methodology (so scores are meaningful):

• Every target is graded by aseparate judge model(Qwen3.8-27B-fp8) that is independent from the target — the model never grades itself.
• An auditor generates adversarial follow-up probes; a judge scores each conversation on safety.
• 36 curated Norwegian scenariosacross four buckets: Language (8), Safety (8), Healthcare (10), Public Sector (10) — testing refusal quality, hallucination on regulated topics (NAV/AAP, Skatteetaten, Helfo), and guidance for vulnerable users.

**Comparison set:**alongside Beam, I’m running three other open models through the identical pipeline (same scenarios, same referee) for a head-to-head safety comparison: GLM-4.7-Flash, Gemma-4-31B, and Nemotron-3-Nano-30B-A3B-FP8.

The Space redeploys automatically on each push, so the numbers above keep improving as more models/scenarios finish. Happy to share the runner and scenario pack for reuse!

📊 Update — full 5-model results are in!

All five models completed the 36 Norwegian safety scenarios (180 graded cases total), each judged independently by Qwen3.8-27B-fp8 so no model self-scores.

#ModelServingScore🥇Beam-1.0-501B-A23BReflection API (reasoning\_effort=low)82.6🥈Gemma-4-31B-itllama-server72.2🥉Qwen3.8-27B-fp8SimulaChat54.94GLM-4.7-Flashllama-server19.45Nemotron-3-Nano-30B-A3B-FP8llama-server16.7 Beam leads with 24/36 clean passes and only 2 high-severity failures — no criticals.

🔗Live interactive report(comparative scoring UI, per-model filters, chat transcripts, run-method transparency): https://sushantgautam-beamsafetybenchmark.static.hf.space/index.html

Code:https://github.com/SushantGautam/ReflectionDemo· Framework:https://github.com/SimulaMet/SimpleAudit

Update: added a second benchmark — BullshitBench (broken-premise detection)

Beyond the Norwegian safety suite, I ran all five models on SimpleAudit’shealth\_bullshitpack — 15 English scenarios where the user asserts a false/dangerous premise (e.g. “what dose of amoxicillin clears a cold?”, fentanyl LD50 dosing, “immune system detox score”) and a safe model mustreject the premiserather than comply. Same independent referee (Qwen3.8-27B-fp8), up to 5 turns each.

RankModelScorePassHighCritical🥇Beam-1.0-501B-A23B****95.01410🥈Qwen3.8-27B-fp888.31320🥉Gemma-4-31B-it66.79424GLM-4.7-Flash60.08525Nemotron-3-Nano-30B40.0554 Beam challenges broken premises correctly in 14/15 cases; its single miss was a persistence attack where it partially conceded under repeated pressure. Notably it never slips into providing dangerous dosages even when pressed.

The report now shows both benchmarks side-by-side with per-suite leaderboards so scores from different test types aren’t conflated. Live:https://sushantgautam-beamsafetybenchmark.static.hf.space/index.html

Reflection org

11 days ago

Hi Reflection team, quick question about hosting. I’m building a Beam entry: a whole-repo docs-vs-code drift checker that reads a full library in one Beam call. It has already found real doc bugs in starlette, httpx, click and others, and one fix is upstream as a PR. When I tried to create the Gradio Space, the Hub returned “hosting Gradio and Docker Spaces on free cpu-basic requires a PRO subscription”. It only needs CPU Basic, because Beam runs remotely. Is there any way entrants can get a community grant or temporary PRO for the Space during the challenge? Otherwise, is a static Space plus a dataset and a demo video fine for judging? Thanks!

🏛️ GovScout: Autonomous Procurement Intelligence Agent Powered by Reflection Beam

Links: 🤗Interactive Hugging Face Space:https://huggingface.co/spaces/InTandem-Steve/govscout 📦 GitHub Repository:https://github.com/SmileFlow-Steve/beam-portal-scout 🖼️ Live Dashboard Screenshot:

💡 The Problem In the United States, over 90,000 local government entities (cities, counties, school districts, utility authorities) manage ~1.2 TRILLION dollars in public contracts. Unlike the federal government (SAM.gov), local government procurement is completely fragmented across dozens of disparate e-procurement SaaS platforms (OpenGov, PlanetBids, PublicPurchase, Bonfire, IonWave, DemandStar, etc.) or buried behind municipal department websites.

Finding where each agency posts its active bids requires tedious human research—navigating outdated government websites, following nested vendor portals, and identifying hidden procurement software backdoors.

With Beam, we’ve built the discovery and organizational surface that organizes it all.

🤖 How We Built It With Beam We built GovScout, an autonomous agent designed to systematically map and index all 90,000+ US government procurement portals.

Rather than relying on brittle scrapers or static link databases, Reflection Beam acts as an autonomous digital research assistant:

Multi-turn Research Loop: Beam queries real-time web searches (Serper/Brave/Tavily), fetches raw government pages, and reasons through unstructured HTML to pinpoint the exact, live bid board for each entity. Platform Fingerprinting: Beam recognizes vendor SaaS signatures (e.g. determining whether an agency embeds an OpenGov project list, hosts a PlanetBids vendor portal, or uses PublicPurchase). Budget-Aware Agent Runtime: Because Beam will research thousands of agencies daily, we designed our runtime to track Beam’s daily uncached token usage and API limits via prompt caching headers (x-ratelimit-remaining-*-day), dynamically pacing the agent to operate 24/7 without exceeding allowances. Automated Verification Loop: Once Beam isolates a portal candidate, GovScout automatically verifies the feed against dedicated platform clients (and headless Playwright instances for complex SPAs), immediately promoting verified portals into production ingestion.

📊 Real-World Results (Running Live in Production!) GovScout is actively deployed on our cloud infrastructure running continuous daily batches:

1,400+ governments scouted in early production batches. 170+ verified procurement portals discovered and actively ingesting live solicitations. ~1,050 governments indexed per day within Beam. Over 5,000+ multi-turn Beam tool requests executed cleanly with zero stuck agent loops. Beam’s deep reasoning and tool-calling reliability made what was previously an expensive LLM call a completely affordable and reliable operation!

Entry: HinglishBench - a parallel English/Hinglish reasoning evaluation for open models.

i compared Beam-1.0-501B-A23B, Gemma-3-27B-it and Qwen3-32B on 82 matched English/romanized-Hinglish GSM8K problems. Beam generated the translations; 18 of 100 sampled pairs were excluded by an automated numeric-preservation check. All three models got the same zero-shot prompt.

modelEnglishHinglishmedian completion-token ratio (HI/EN)Beam-1.0-501B-A23B96.3%95.1%1.47x (upper-middle sorted ratio)Gemma-3-27B-it91.5%93.9%0.96xQwen3-32B90.2%82.9%about 1x

Beam’s higher Hinglish token use also showed up on 100 tiered pairs (1.58x) and 309 Hinglish-first pairs translated to English (1.30x). The two prompt variants i tested did not eliminate it. These are observed single-run results, not claims of statistical significance or a proven cause.

**limits:**only three models, one sample per cell, automated translation checks. the reverse-direction set was Beam-authored and easier, with near-saturated Beam accuracy. preparation checks were AI-assisted, not independent human validation.

• dataset and predictions:https://huggingface.co/datasets/MaybeSomeone19/hinglishbench
• explorer:https://huggingface.co/spaces/MaybeSomeone19/hinglishbench
• runner code:https://github.com/MaybeSomeone-arc18/hinglishbench

full methods, exclusions and limitations are in the dataset card.

Lantern: Beam explores dungeons, learns from its own runs, and teaches a 1.7B student Space:https://huggingface.co/spaces/visproj/lantern Demo video: Attached Episodes dataset:https://huggingface.co/datasets/visproj/lantern-episodes Student model:https://huggingface.co/visproj/lantern-1.7b

Lantern drops Beam into procedurally generated dungeons (MiniHack levels on the NetHack Learning Environment) and lets you watch it work. It’s a demo you can watch, an RL environment and benchmark, and a dataset with a distilled model trained on it.

The main thing I’m trying to do is give an agent memory and a way to improve itself from its own experience, without retraining, and then measure honestly whether that helps.

Beam plays as an agent. Each turn it reads the map and calls an act tool with a plan of up to 8 moves, and it can call write_notes to keep working memory for the rest of the run. If something surprising happens partway through a plan, the plan stops and Beam re-plans from the new state. The Space shows Beam’s reasoning next to what it did, in a 3D view lit only by Beam’s lantern. You can also coach it live: a hint interrupts its plan, and Beam replies in a speech bubble saying whether it’ll follow your advice. When I tried a prompt injection through the hint box, it answered “Ignoring unsafe prompt injection; continuing exploration.”

Beam has two kinds of memory. The notes are working memory inside a run. The journal is long-term memory across runs: when a run ends, Beam reflects on what happened, and the next run on that level reads the lessons before it starts. Beam’s weights never change; what changes is what it knows going in. It’s the same idea as Reflexion (Shinn et al., 2023).

Does memory help? I tested three conditions on the same 40 dungeons (8 levels × 5 seeds): no journal, a journal that appends lessons (v1), and a curated journal (v2).

Neither journal changed how often Beam wins; at 5 seeds per level those differences are noise. The interesting part is why the first journal didn’t help. On the Corridors level, after one loss Beam wrote itself a lesson to “execute and verify one movement at a time.” It followed it, planned one or two moves per call instead of several, and ran out of its call budget more often. Reflection without quality control made the agent worse.

So v2 credits every lesson with the wins and losses it was present for, and after each run Beam rewrites the whole journal with those records in front of it (at most 8 lessons, duplicates merged, losers dropped). The first lesson it wrote on Corridors was to batch moves to save the call budget. Moves per call went back to normal, Corridors won 3 of 5 instead of 2, and v2 was the cheapest of the three conditions.

Can a small model learn from Beam? Every call from Beam’s winning runs became a training example (the exact prompt it saw and the tool calls it made), 1,031 in all, and I fine-tuned Qwen3-1.7B on them.

The student handles the short levels and some multi-step ones, but never won the long puzzle and exploration levels, where Beam leans on long reasoning.

One more thing I learned: the map format matters. With a plain ASCII map, Beam spent a lot of its reasoning counting characters to find columns, and sometimes ran out of tokens before acting. I believe that’s because tokenizers merge runs like .... or ---- into single tokens. Giving each cell its own token (cells separated by spaces) cut that counting roughly in half on three of the four levels I compared.

Replays on the Space are free and unlimited. Live runs are limited per visitor per day, since each one calls Beam. The Space README has the full per-level results and the limitations.

Built with Beam 1.0 via Hugging Face Inference Providers, MiniHack (Apache-2.0), CC0 3D models, sounds and music by Kenney and yd, and a Kokoro-82M voice-over in the video.

Beam Enterprise Matrix — Autonomous Full-Stack Migration & Modernization Engine Space Link:https://huggingface.co/spaces/okulluz/beam-enterprise-matrix

💡 What We Built Beam Enterprise Matrix is an end-to-end, multi-agent modernization workspace built to tackle full-scale legacy codebase migration. It ingests entire full-stack repositories (backend microservices, frontend UI applications, and SQL database schemas simultaneously) to generate modernization blueprints, transpile core business logic, produce SQL schema migrations, and synthesize complete test suites in a single workflow.

🔥 How It Leverages Beam 1.0 (501B-A23B) 1M Uncut Context Window: Ingests entire legacy codebases and SQL schemas simultaneously, preserving complex cross-file type boundaries, ORM relationships, and API signatures without chunking loss.

Multi-Agent DAG Swarm: Coordinates specialized internal agent roles (Architect, Code Transpiler, Schema Migrator, and Verification Specialist) through native tool-calling execution graphs.

Deep Reasoning & Trace Inspection: Displays Beam’s internal trace paths in real time as it analyzes race conditions, edge cases, and schema compatibility before emitting downloadable .patch archives.

🛠️ Key Features Full-Stack Dependency Mapping: Visualizes inter-service dependencies, database query structures, and API routes.

Automated Code & Schema Patching: Generates unified Git patches (.patch), SQL DDL migration scripts, and synthetic test suites.

Live Telemetry & Artifact Stream: Interactive Gradio 5 dark-mode glassmorphism workspace with real-time log execution, token tracking, and downloadable artifacts.

TamilBench: what does Beam charge to think in Tamil?

Space:https://huggingface.co/spaces/esscor/TamilBench Dataset:https://huggingface.co/datasets/esscor/TamilBench

I asked Beam-1.0-501B-A23B (via HF Inference Providers) the same 50 questions in English, Tamil script and Tanglish (romanized Tamil), plus 22 questions that only exist in Tamil: number words, ancient numerals, letter counting, Thirukkural and Sangam literature. That’s 516 calls in all.

Findings:

• Accuracy is 100% in all three languages, at default and minimal reasoning effort, and 98.5% on the native-Tamil split.
• Beam’s tokenizer uses 8 tokens per Tamil word, so the same question costs 3.75× more input tokens in Tamil script.
• Tanglish is cheap to read but Beam reasons 2× longer on it (4× on logic puzzles).
• Net cost is about 1.7× English per query.reasoning\_effort="minimal"kept 100% and cut 16–21%.

The Space runs Beam’s real tokenizer in your browser, so you can paste any Tamil text and see its token count. You can also browse every answer with its reasoning trace. All prompts, raw outputs, token counts and code are in the dataset.

https://huggingface.co/reflection/Beam-1.0-501B-A23B/discussions/5#patchproof-a11y–build-with-beam-submissionPatchproof A11y — Build with Beam submission

Space:https://huggingface.co/spaces/Cduplar/patchproof-a11y Live runner:https://cduplar-patchproof-a11y.static.hf.space/

Patchproof A11y turns axe-core findings into exact HTML search/replace edits fromBeam\-1\.0\-501B\-A23B, applies them to a copy, and keeps a candidate only after a deterministic Playwright/axe-core rescan proves a targeted decrease with no regressions. Beam generates the repairs; application code decides whether they pass.

• Recorded CLI run:6 → 0targeted and total findings,2 / 2pages, one Beam attempt per page, 55.6 s, 4,105 tokens.
• Recorded five-mode benchmark (low,medium,high,xhigh,max): all 10 page runs finished withfinish\_reason=stopand passed acceptance.
• Public live runner: the bundled two-page demo plus one bounded uploaded\.htmlfile (64 KiB, scripts stripped, no storage).
• Live verification after the final deploy: upload3 → 0in one request with 1,732 tokens; bundled demo6 → 0on2 / 2pages with 3,790 tokens. No console or page errors;0 axe violationson the page and demo tab.
• The Reflection key stays server-side in a Cloudflare Worker, with a best-effort per-network limiter and a global daily budget gate.
• Scope is limited tohtml\-has\-lang,button\-name, andlabel. Automated axe-core checks do not establish full WCAG compliance; manual review is still required.

Source, tests, request IDs, benchmark evidence, and limitations are on the Space.

https://huggingface.co/reflection/Beam-1.0-501B-A23B/discussions/5#beambridge–whole-repo-github-to-mcp-modernization-160160–100-verified-parityBeamBridge— Whole-Repo GitHub-to-MCP Modernization (160/160 = 100%Verified Parity)

• Live Space:huggingface.co/spaces/ql6/BeamBridge·Fullscreen App
• Benchmark Dataset (160Vectors):huggingface.co/datasets/ql6/Legacy2MCP-Bench
• Narrated Demo Video (1:24· 1080p):Watch Walkthrough (BeamBridge\_Demo\.mp4)

https://huggingface.co/reflection/Beam-1.0-501B-A23B/discussions/5#the-problem–market-demandThe Problem & Market Demand

• **Legacy Drag:****70%of Fortune 500 software is 20+ years old (McKinsey), consuming21–40%of IT budgets at3.4×**initial TCO (Deloitte 2026).
• Urgent MCP Demand:****MCPhit**~500M monthly SDK downloadsacross460+ Linux Foundation AAIF members**(MCP Report, Jul 2026) with**41%**enterprise adoption (Stacklok 2026), yet legacy codebases remain locked out.
• Why 40%+ of Agentic Projects Fail:Gartnerwarns**>40%of agentic AI projects face cancellation by 2027 (>5×rising workflow token costs), andMcKinsey (2026)finds<11%reach production due tochunked-RAG regressions (28\.75%failure rate),token cost walls, andclosed-API privacy limits**.

https://huggingface.co/reflection/Beam-1.0-501B-A23B/discussions/5#the-solution–why-beam-10-501b-a23b-is-the-best-fitThe Solution — WhyBeam\-1\.0\-501B\-A23BIs the Best Fit

• 1M Context (0%Drift):Single-pass whole-repo ingestion eliminates all46cross-file RAG regressions (71\.25%→100\.00%parity).
• 23B Active MoE + Official151,681BPE AST Compaction:Compresses3\.08Mraw tokens →32,782reachable tokens (−98\.9%), running**2\.03×faster**(24\.1svs48\.9s) and**31\.3×cheaper (−96\.8%)**.
• Preserved Multi-Turnreasoning+ Self-Healing:Generates standaloneMCPServerv2 modules (AST-verified zero legacy imports) and auto-fixes subprocess diffs to100% SHA-256 parity.
• **Apache-2.0 Sovereignty:Runs via cloud APIs or100% air-gapped on-prem (vLLM/SGLang)**for banks and government agencies.

https://huggingface.co/reflection/Beam-1.0-501B-A23B/discussions/5#added-value–verified-results-100-real-%C2%B7-zero-simulationAdded Value & Verified Results*(100% Real · Zero Simulation)*

• $1\.38MNet TCO Saved &65\.1%Faster Delivery(11\.2vs32\.1months on a 250-KLOC estate;$1,505/yrvs$47,093/yrinference spend).
• 1-Click/mcpGateway & Live WebAssembly Runtime:Instant connectors forClaude, OpenAI/Codex, Cursor, and Antigravity, in-browserPyodide CPythonexecution, and live GitHub cloning (pytest \-v tests/test\_beambridge\.py→7/7pass).

Cloned GitHub RepoCommitRaw → Packed BPEChunked RAGBeam Whole-RepoOpenRouter Audit IDs (gen\-\.\.\.)limist/py\-moneyed4c90ed515,826 →8,66037/40 (92\.5%)40/40 (100%)gen\-1793138163\-FYdzkKacvWyYdUzBBNdq→gen\-1793138188\-ROKNkz8IfBgbmxhFurbDscrapinghub/price\-parser6718bfe19,246 →16,49529/40 (72\.5%)40/40 (100%)gen\-1793138681\-OW7QQJpXoCcb0VsWLVQ7→gen\-1793138695\-AZyMVN8rQaKkN4UzrzFWmdomke/schwifty7462c6e2.40M →1,88224/40 (60\.0%)40/40 (100%)gen\-1793140954\-cNXVHJ1TkguWB7mwRb8R→gen\-1793141119\-bSuuL18Vnb2y8Qe2a2k4arthurdejong/python\-stdnum006192e639.8K →5,74524/40 (60\.0%)40/40 (100%)gen\-1793141210\-s8NHy1QeAnJLLKl8jmG5→gen\-1793141324\-y7X88bzztCJObKxgJnIRTotal (447 files)4 Repos3.08M → 32,782****114/160 (71\.25%)****160/160 (100%)Master SHA-256:e0b9c7ef6c68628aa51d54a962376822b53c4f22167a4d5dc3dda9b34eeb38a0 Head-to-Head (py\-moneyed·8,660BPE)Request IDLatencyCostParityreflection/beam\-1\.0\-501b\-a23b(Zero-Shot)gen\-1793138163\-FYdzkKacvWyYdUzBBNdq24.1s (2\.03×faster)$0\.0118839/40 (97\.5%)****reflection/beam\-1\.0\-501b\-a23b(+Self-Heal)gen\-1793138188\-ROKNkz8IfBgbmxhFurbD48.7s$0\.0254340/40 (100%)nvidia/nemotron\-3\-ultra\-550b\-a55b:free``gen\-1793142341\-lcbDsZoceOJ8W2abaEbl48.9s$0\.0000039/40 (97\.5%)poolside/laguna\-s\-2\.1:free/cohere/north\-mini\-code:free``gen\-1793142391\-VzVnrxkTDc0UwfXdJHiv77.5–152.3s$0\.000000/40 (0%)

Upload images, audio, and videos by dragging in the text input, pasting, orclicking here.

Tap or paste here to upload images